AI Can Help Write Your EU Proposal. It Cannot Make Your Consortium Credible.

Fluency used to be a rough signal of effort. A team that wrote well had usually thought hard. Generative AI has broken that signal.

Text can now sound credible without being true. And every claim of access, capacity or commitment will be read by an evaluator looking for the partner, task and effort behind it.

The Commission’s guidance for Horizon Europe allows applicants to use generative AI when preparing proposals. Applicants must be transparent about the tools used and remain fully responsible for the content, including parts an AI produced. Using AI is not, in itself, grounds for evaluators to penalise a proposal.

Where AI earns its place

Used well, AI strengthens both drafting and analysis. It can:

• summarise a long call text and extract its requirements;
• propose a structure that follows the template;
• check consistency between sections, such as objectives against indicators;
• tighten language and remove repetition;
• identify gaps, unsupported claims and missing explanations;
• argue an evaluator’s case against your draft.

The last item is the one most teams skip. Ask a model to make the strongest case against your Impact section and you will often get a sharper list than a polite colleague provides. AI can support strategic judgement, not just tidy prose.

What no model can do for you

Some parts of a proposal can only be produced in the real world. AI cannot recruit the right partners. It cannot agree responsibilities between organisations. It cannot validate that users actually need what you plan to build. It cannot secure access to a hospital, a factory or a school. And it cannot obtain a genuine commitment from anyone.

Those are not writing problems. They are delivery problems that happen to show up in writing.

Fluent is not the same as true

A hypothetical example. An AI-assisted draft contains this sentence:

“Through established partnerships with three regional hospitals, the consortium has secured pilot sites and access to anonymised care-pathway data for 2,400 patients, enabling robust validation in real clinical settings.”

It reads well. It sounds specific. It is also unsupported. No hospital is a partner. No access agreement exists. Nobody has discussed data protection with anyone.

An evaluator will look for the hospitals in the consortium, or for evidence of an access agreement, and for the tasks and effort behind the pilots. Finding none, they will treat the claim as unsubstantiated. Worse, they may start doubting the claims around it.

The sentence did not fail because AI wrote it. It failed because nobody checked whether it was true.

Dividing the work between AI and people

A practical division looks like this:

• AI drafts, structures and checks. Use it to produce first versions, compare sections and flag inconsistencies.
• People decide and verify. A named person confirms every factual claim, number and commitment.
• Partners confirm their own commitments. Each partner signs off the description of their role, effort and resources.
• A reviewer challenges the whole. Someone outside the drafting process reads it against the criteria.

One useful habit is a claims register: a simple list of every material claim, its evidence and who verified it. A claim without evidence is either removed or turned into planned work.

Handle confidential material responsibly

Proposals contain unpublished methods, budgets, personal data and partners’ commercial information. Before pasting any of it into an AI tool, check how that tool stores and uses inputs, and whether your consortium has agreed to it.

Partners who shared information in confidence did not necessarily agree to share it with a third-party service. Ask first.

Bottom line. AI can help you articulate and assess your evidence. It cannot create it. Institutional commitment, delivery experience, validated results and partner capacity have to exist before any tool can describe them.

Audit your current draft: highlight every claim about access, capacity, results or commitment. Then mark each one as either evidenced or merely convincing. The second list is your real to-do list.